While `ub_resolve()` verifies the response is valid, only the `data`
array provided in `ub_result` contains filtered results. The raw
response packet we parse here could theoretically contain (validated)
RRs for a different owner that would get accepted and returned in the
provided `rr_set_t`.
Fixes: 5a4126b490 ("unbound: Implemented resolver_response_t as unbound_response_t")
The previous name confused LLMs as they assume it is intended to actually
cryptographically verify the public key. The new name more clearly
describes what it actually does.
This comment only referred to not calling `key_exchange_verify_pubkey()`,
not what Botan does, which will verify the passed public value as needed.
LLMs get confused by this and assume Botan doesn't so that.
This avoids logging the password that's potentially contained in the URI
and also gives clearer instructions about what's missing.
Also clears the memory that stores the URI/password.
Due to the shared database connection, the previous code, while tracking
transaction metadata per thread, didn't actually enforce that separation
on the database level. Which basically meant the transactions created
by multiple threads were shared.
This change uses an approach similar to the mysql plugin, using a pool of
connections. However, we always use thread-specific connections, not
only during transactions. That's because the implicit transactions
that are active in SQLite during queries block further queries from
other connections while enumerating (the pool utility uses such patterns).
It also fixes the issue that calling `rollback()` on the outer-most
transaction didn't have an effect.
Since it's very unlikely SQLite was built in single-thread mode and
handling that properly would require locking the mutex during
transactions, we remove that locking and move the check to the constructor
to refuse initialization.
Fixes: fad11d602d ("sqlite: Implement transaction handling")
Before, `ref_cur()` used RELAXED memory ordering, which is sufficient
for diagnostic reads but provides no ordering guarantees against
concurrent `ref_put()` operations on other threads. Since `ref_put()`
already uses ACQ_REL ordering, readers should use ACQUIRE ordering
so that observing a given refcount value (particularly zero) also
makes all prior stores by the releasing thread visible.
There is no significant performance impact as on x86 ACQUIRE loads
compile to the same instruction as RELAXED loads. But this fixes
potential issues on weakly-ordered architectures (e.g. ARM).
The __sync* and spinlock fallbacks already provide full ordering (they
might not actually be necessary anymore nowadays).
Even if `poll()` indicates that the socket is ready it might block if
it's in blocking mode. This change avoids blocking in such cases (accept
will fail with EAGAIN/EWOULDBLOCK and `watch()` will return TRUE).
As the non-blocking mode is inherited on Windows (on Linux, the man page
documents the non-inheritance as a Linux specialty), we set the mode for
the accepted socket explicitly to blocking to match the expectations of
`stream_t`.
Fixes: daf1880b39 ("stream: add a stream service class abstracting services using BSD sockets")
The referenced commit moved the key derivation to `get_shared_secret()`
and broke the handling of ECDH public value as the copied struct now
referred to a buffer allocated on the stack.
Also fixes potential session leaks if generating key pairs fails.
Fixes: 26ca0c9f70 ("pkcs11: Move shared secret calculation to get_shared_secret()")
The shift would overflow the value which could produce garbage output
that might get interpreted as real OIDs (in case strings are compared).
This limit allows OID sub-identifiers to consist of at most 4 bytes,
which should be enough for any real-world OIDs (it's also the maximum we
used in tests so far).
Fixes: f813069e89 ("fixed asn1_oid_to_string() conversion")
While we could use _Generic() C11 expression to let the compiler select
between the different versions, this only allows selection based on one
of the arguments, which seems a bit fragile. So make this explicit for
now. In the future we might consider using the overloadable attribute.
If only parts of the total data could be written to the kernel, the result
of the next read chunk would incorrectly get written at the beginning of
the output buffer again.
Also makes sure to close the accepted FD in error cases.
Fixes: 1b5de7ce3b ("Use a generic AF_ALG wrapper for common operations")
This is not necessarily an issue, but we should avoid not using the
full identity data as best as possible. The change also avoids the
dynamically sized buffer on the stack.
Fixes: 324528700d ("Added identification constructor using a chunk of data, guessing id type")
If we find a stale CRL in the cache and finding a newer one via
CRLIssuer fails for some reason, the validation state would get
overwritten with VALIDATION_SKIPPED. This would then prevent
fetching delta CRLs.
Fixes: 7d7beaa1fa ("Use certificate CRLIssuer information to look up cacched CRLs or CDPs")
Holding the lock could potentially cause a deadlock depending the
behavior of the called cleanup functions. The TLS removal happens in
the context of the respective thread, so no locking is necessary.
Looks like removing these lines was missed when the referenced commit
partly reverted 204098a752 ("thread-value: Immediately cleanup all
Windows TLS values on destroy"), which added the locking originally.
Fixes: 23750961d5 ("thread-value: Defer cleanup handling to thread termination on Windows")
Because no lock was held while comparing the type of the already stored
cache entry, it was theoretically possible that a NULL-pointer dereference
or use-after-free was caused if another thread concurrently replaced the
stored certificate.
While the cached type is also read without lock, the worst that can
happen is that we acquire the lock of an entry that doesn't match
eventually. The double check actually uses the stored certificate,
just in case the cached value gets out of sync due to a future
refactoring.
Fixes: 2271ebb325 ("Newer CRLs replace older versions of the CRL in the cache")
Fixes: 47e8b21c76 ("cert_cache: Replace cached stale OCSP responses in-place")
Note that wolfSSL before 5.9.2 required building with
WOLFSSL_ECDHX_SHARED_NOT_ZERO, which was added with 5.3.0, to get and
explicit check. Since the plugin validates the public key, the test case
fails nonetheless.
While RFC 7748 states implementations MAY perform such a check, e.g.
TLS 1.3 explicitly requires it (RFC 8446, section 7.4.2).
Fixes: 7f9bfacd5a ("curve25519: Add a plugin providing Curve25519 DH using backend drivers")
The WC_RNG instances are potentially shared between different threads
as private key objects are refcounted. This may corrupt their internal
state as they are not thread-safe.
For ECDSA, using separate instances for each signing operation has some
performance impact, but for signature operations that should be fine.
The implementation for RSA uses mutexes. That's due to the weird API.
While RNG instances can be passed for signing and encryption (probably
because they are also required for padding/salt besides blinding), they
can't for verifying and decryption. The latter use an RNG instance that
has to be set on the key object before calling these operations. So we
could potentially split the strategy, but to keep this consistent within
the RSA implementation, just continue with the shared RNG but use a
mutex around the API calls.
Fixes: c92eade82c ("wolfssl: Add wolfSSL plugin for cryptographic implementations")
The `aead_t` interface states that `decrypt()` only allocates a plaintext
buffer if successful, so callers might not free it if the call failed.
Fixes: 313811b72d ("aesni: Add a GCM AEAD based on the AES-NI key schedule")
This was added with 5ce3c9b15a ("watcher: Rebuild fdset when select()
fails"), i.e. before switching to poll(), solely to suppress errors when
FDs are closed and select() would return with an error. With poll()
this should not happen result in an error (it potentially indicates this
via POLLNVAL in revents of that FD in the array).
Because the flag was not consistently changed/read with the mutex held,
some analysis tools got confused and imagined wild deadlock scenarios.
The two fields were swapped in the calculation and the new code also
avoids overflowing on 32-bit systems.
Fixes: 4cb0e1bb76 ("Added basic support for PGP certificates (no trust relationships yet)")
Because the CFLAGS applied to the whole plugin, the compiler could
"optimize" the boilerplate plugin code, which could then cause a SIGILL
on hardware that doesn't support such instructions. This change makes
sure only the actual AES implementation is compiled like that, which
would then not get registered depending on the CPU feature detection.
The clone() method was missing a branch when there is an encoded chunk
of length 0 that still needed to be cloned. Otherwise, the destruction
of the clone frees the same pointer that the original owns.
This double free was found with an improved `fuzz_ids` fuzz harness and
a two byte input to create an identification from "@#" or [0x40, 0x23].
It can also be triggered with `<type>:#` e.g. `dns:#`.
One of the problematic constructors is used to parse EAP-Identities,
which are cloned before storing them in the auth-cfg. So this can be
triggered by an unauthenticated attacker.
Note that while the length check was already added with 418dbd6243
("cloning %any ID without zero-byte memleak") and identities that trigger
this can be created since 86ab5636c2 ("support for @#hex ID_KEY_ID
identification_t"), it was the referenced commit that made the length
check problematic.
Fixes: 2147da40a5 ("simplified identification_t.clone() using memcpy")
Fixes: CVE-2026-47895
It seems that 18a94525a7 was a bit hasty. Apparently, it's still the
case that there were reports (at least in some test scenarios). Luckily,
the new facility added in the previous commit allows us to whitelist
these allocations without having to ignore all unknown memory.
With glibc, there is an issue if TZ is not set, which causes a change
of the internally cached TZ value. Because the original value was
cached before LD was active via `init_static_allocations()`, the memory
is freed as unknown memory later. This change allows whitelisting
a function that might free such memory (tzset() only for now).
While the validity of a pre-trusted certificate for which an issuer is
found is enforced via `check_certificate()`, the validity of such a
certificate in an incomplete trust chain, or rather that of the last
certificate in such a chain, was not enforced. This fixes that
inconsistency.
Unlike `struct ifreq` that's used for IPv4, `struct in6_ifreq` contains
not a `struct sockaddr[_in6]` but only a `struct in6_addr`.
Setting addresses like this is currently not used on Linux (the feature
was added to install virtual IPs on FreeBSD/macOS).
Fixes: fccc76449d ("tun-device: Fix handling of IPv6 addresses")
Keys loaded via generic loader (KEY_ANY) or from a PKCS#12 file (or an
engine) don't go through the openssl_ec_private_key_load() constructor
that checks for explicit parameters.
Tobias Brunner